A power energy consumption analysis method, system, device and medium based on multi-source data fusion

Abstract

This invention discloses a method, system, device, and medium for power energy consumption analysis based on multi-source data fusion, belonging to the field of power energy consumption analysis technology. The specific steps are as follows: Historical and real-time multi-source working data are collected and preprocessed; behavioral event feature vectors are extracted from a multi-layer multi-source aligned dataset; a multi-level energy consumption and carbon emission baseline dataset is established using a graph neural network; multi-time-domain energy consumption and carbon emission prediction is performed; based on the prediction results, energy-saving and carbon optimization strategies are output and verified in a simulation model; the verified energy-saving and carbon optimization strategies are deployed and executed, and the strategy generation process is optimized based on execution feedback. This invention achieves collaborative optimization management of energy consumption and carbon emissions in complex industrial parks, improves state characterization accuracy through multi-source fusion and behavior-driven approaches, achieves risk identification through multi-time-domain prediction, and ensures the safety and adaptability of the strategies through closed-loop verification and dynamic adjustment.

Description

Technical Field

[0001] This invention relates to the field of power energy consumption analysis technology, and in particular to a power energy consumption analysis method, system, equipment and medium based on multi-source data fusion. Background Technology

[0002] As old industrial parks are transformed towards intelligent and green low-carbon development, they often contain multiple discrete manufacturing production lines, multiple office and dormitory buildings, centralized air conditioning chillers, distributed photovoltaic systems, electrochemical energy storage devices, and a large number of electric vehicle charging stations, forming a complex energy consumption pattern integrating source, grid, load, and storage. Traditional power consumption analysis and energy management systems rely heavily on meter readings and a limited number of equipment operating parameters, primarily focusing on load statistics and simple peak shaving and valley filling control from the power side perspective. This approach fails to adequately consider multi-dimensional factors such as production scheduling, process recipe switching, personnel entry and exit patterns, building environment comfort, multi-energy synergy, and carbon quota constraints.

[0003] Existing methods generally model production process systems, building environment systems, and distributed energy systems separately, lacking a multi-source heterogeneous data alignment mechanism based on a unified spatiotemporal benchmark. This makes it difficult to establish clear data transfer relationships between equipment, region, and park levels, and to characterize the dynamic impacts of behavioral events such as shift changes, work order insertions, peak work hours, and charging peaks on energy consumption and carbon emissions. While some studies have introduced machine learning models for load forecasting or energy consumption assessment, these often remain at the single time series level, with model inputs limited to historical power curves or a few environmental variables. This lack of systematic characterization of production conditions, user behavior, and carbon quota status leads to significant discrepancies between predicted results and actual operation. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: how to construct an intelligent analysis and decision-making system that can achieve multi-level coordinated optimization of energy consumption and carbon emissions in the complex scenario of integrated source-grid-load-storage in old industrial parks, through the fusion of multi-source heterogeneous data with unified spatiotemporal benchmarks and behavior event-driven analysis.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a power consumption analysis method based on multi-source data fusion, comprising, Collect historical and real-time multi-source working data and preprocess them to output a multi-layer multi-source aligned dataset. Extract behavioral event feature vectors from multi-layer, multi-source aligned datasets; A multi-level baseline dataset for energy consumption and carbon emissions is established by combining behavioral event feature vectors with graph neural networks. Based on the multi-level energy consumption and carbon emission baseline dataset, multi-time domain energy consumption and carbon emission predictions are performed, and the multi-time domain energy consumption and carbon emission prediction results are output. Based on the multi-time-domain energy consumption and carbon emission prediction results, a reinforcement learning algorithm is used to output energy-saving and carbon optimization strategies and verify them in a simulation model. The validated energy-saving and carbon optimization strategies are deployed for execution, and the strategy generation process is optimized based on the execution feedback.

[0007] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the step of extracting behavioral event feature vectors from a multi-layer multi-source aligned dataset includes: Identify behavioral events in multi-layer, multi-source aligned datasets to form behavioral event samples; Based on behavioral event samples, construct behavioral event feature vectors to form a feature library; A causal prior matrix is ​​established based on the behavioral event feature vectors in the feature library.

[0008] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the step of establishing a multi-level energy consumption and carbon emission baseline dataset based on behavioral event feature vectors combined with graph neural networks includes: A hierarchical heterogeneous energy consumption map is constructed based on a multi-layer, multi-source aligned dataset and feature library. The data in the multi-layer multi-source aligned dataset is mapped to the features of each node in the hierarchical heterogeneous energy consumption graph, and the feature library and causal prior matrix are mapped to the features of the edges and the initial values ​​of the edge weights in the heterogeneous energy consumption graph. The mapped hierarchical heterogeneous energy consumption map is input into the graph neural network fusion model to obtain the embedding features of each node; Based on the embedding features of each node, a multi-level energy consumption and carbon emission baseline dataset is generated under given operating conditions and behavioral patterns by combining a multi-level, multi-source aligned dataset.

[0009] This invention integrates multi-source data and behavioral event features, and utilizes graph neural networks to construct multi-level energy consumption and carbon emission baselines that reflect the actual operating state. The established baselines are dynamically adaptable and can accurately characterize the energy consumption and carbon emission patterns under specific operating conditions and behavioral modes, achieving a refined and hierarchical portrayal of the system's operating state.

[0010] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the step of performing multi-time-domain energy consumption and carbon emission prediction based on multi-level energy consumption and carbon emission baseline datasets, and outputting multi-time-domain energy consumption and carbon emission prediction results, includes: Based on the park's operational needs, short-term, medium-term, and long-term forecast time domains are defined, and a correspondence is established between the multi-level energy consumption and carbon emission baseline datasets and the multi-level, multi-source aligned datasets. The data from the multi-layer, multi-source aligned dataset and the multi-level energy consumption and carbon emission baseline dataset are combined into a multi-time-domain prediction input feature sequence. A graph-driven time-series prediction model is established, which takes multi-time-domain prediction feature sequences as input and outputs multi-time-domain energy consumption and carbon emission prediction results.

[0011] This invention achieves multi-time-domain collaborative prediction of energy consumption and carbon emissions from equipment to the entire industrial park by integrating multi-level baselines that combine behavioral and physical characteristics with information from the external environment and production plans, and using a graph-driven time-series model. The prediction results not only generate power and carbon emission curves for future periods, but also simultaneously quantify multi-dimensional operational risks such as transformer load, indoor comfort, process quality, and carbon quotas.

[0012] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the step of using a reinforcement learning algorithm to output energy-saving and carbon optimization strategies based on multi-time-domain energy consumption and carbon emission prediction results and verifying them in a simulation model includes: An expert rule base was established based on safety procedures and operation and maintenance experience, and the preset rules were transformed into action constraints and penalty parameters for energy-saving and carbon optimization strategies. The action space is defined by using multi-time-domain energy consumption and carbon emission prediction results, multi-level energy consumption and carbon emission baseline datasets, and feature libraries as state inputs to the reinforcement learning algorithm. Establish a reward function based on the penalty parameters; The reinforcement learning algorithm searches the policy space to generate a candidate set of energy-saving and carbon optimization strategies that meet the constraints of the expert rule base and the preset reward value range. The candidate set of energy-saving and carbon optimization strategies is input into the simulation model for verification.

[0013] This invention combines multi-time-domain prediction results with an expert rule base and performs collaborative search in the action space through reinforcement learning. It utilizes an electrothermal behavior coupled digital twin model to perform perturbation simulation and security verification on candidate strategies, ultimately outputting a reliable and executable energy-saving and carbon optimization strategy that simultaneously optimizes electricity costs, carbon emissions, comfort, and equipment lifespan while satisfying various hard constraints.

[0014] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the step of inputting the candidate set of energy-saving and carbon optimization strategies into the simulation model for verification includes: Input the candidate set of energy-saving and carbon optimization strategies into the digital twin model of electrothermal behavior; The electrothermal behavior coupled digital twin model constructs a coupled simulation structure of the park power grid, building environment, and production process based on a multi-layer multi-source aligned dataset and a multi-level energy consumption and carbon emission baseline dataset. Each strategy in the candidate set of energy-saving and carbon optimization strategies is regarded as a perturbation to the current baseline operating conditions. The perturbation is applied to equipment-level, regional-level, and park-level nodes to simulate power operation data, building environment data, distributed energy and energy storage data, and carbon emission changes after strategy implementation. The model compares energy consumption and carbon emission indicators and risk indicators before and after strategy implementation. Strategies that violate expert rule base constraints or cause risk indicators to exceed limits are eliminated. The remaining strategies are sorted according to their comprehensive benefit scores, and the energy-saving and carbon optimization strategies are output.

[0015] As a preferred embodiment of the power consumption analysis method based on multi-source data fusion described in this invention, the process of issuing and executing the verified energy-saving and carbon optimization strategies and generating optimization strategies based on execution feedback includes: The energy-saving and carbon optimization strategies output by the electrothermal behavior coupled digital twin model are sent to the park control terminal to jointly control the production process control system, building environment control system, distributed energy and energy storage control system, and charging control system. During the strategy execution process, power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data are collected to form a strategy execution feedback dataset. The strategy execution feedback dataset is compared with the multi-level energy consumption and carbon emission baseline dataset and the multi-time domain energy consumption and carbon emission prediction result set. The deviation between the actual energy saving benefit and the actual carbon emission reduction benefit and the predicted value is calculated. According to the source of the deviation, the node features and edge weights in the hierarchical heterogeneous energy consumption graph are updated respectively. The parameters of the graph neural network fusion model and the graph structure-driven time series prediction model are updated. The strategy execution feedback dataset is used as the experience replay data of the reinforcement learning algorithm to incrementally train the energy saving and carbon optimization strategy, forming a strategy effect profile. The strategy effect profile evolves over time.

[0016] This invention provides a power consumption analysis system based on multi-source data fusion.

[0017] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a power consumption analysis system based on multi-source data fusion, comprising: an acquisition module, a feature extraction module, a baseline establishment module, a prediction module, a strategy output module, and a strategy optimization module; The acquisition module collects historical multi-source working data and real-time multi-source working data, performs preprocessing, and outputs a multi-layer multi-source aligned dataset. The feature extraction module extracts behavioral event feature vectors based on a multi-layer, multi-source aligned dataset. The baseline establishment module establishes a multi-level energy consumption and carbon emission baseline dataset based on behavioral event feature vectors combined with graph neural networks. The prediction module performs multi-time-domain energy consumption and carbon emission prediction based on multi-level energy consumption and carbon emission baseline datasets, and outputs multi-time-domain energy consumption and carbon emission prediction results. The strategy output module outputs energy-saving and carbon optimization strategies based on multi-time-domain energy consumption and carbon emission prediction results using reinforcement learning algorithms, and verifies them in a simulation model. The strategy optimization module is the process of issuing and executing the verified energy-saving and carbon optimization strategies and optimizing the strategy generation process based on the execution feedback.

[0018] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the power consumption analysis method based on multi-source data fusion.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the power consumption analysis method based on multi-source data fusion.

[0020] The beneficial effects of this invention are as follows: This invention constructs a multi-layer, multi-source aligned dataset and feature library, utilizes hierarchical heterogeneous graph neural networks and multi-time domain prediction to obtain equipment-level to park-level energy consumption and carbon emission baselines and prediction results, combines expert rule-constrained reinforcement learning and digital twin simulation to generate executable energy-saving and carbon optimization strategies, and dynamically adjusts optimization weights through carbon quota status to achieve collaborative optimization management of energy consumption and carbon emissions in the integrated source-grid-load-storage scenario of old industrial parks. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process of a power consumption analysis method based on multi-source data fusion according to an embodiment of the present invention. Detailed Implementation

[0023] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0024] Example 1, referring to Figure 1 As one embodiment of the present invention, this embodiment provides a power consumption analysis method based on multi-source data fusion, comprising: It should be noted that as old industrial parks evolve into complex energy consumption scenarios integrating power generation, grid, load, and storage, they simultaneously contain multiple loads such as discrete manufacturing, building environment, distributed energy, and electric vehicle charging. The energy consumption pattern exhibits significant multi-source heterogeneity and strong behavioral coupling characteristics. Traditional energy consumption analysis methods often rely on single meter data, making it difficult to achieve unified alignment and fusion analysis of multi-dimensional data such as power operation, production processes, building environment, user behavior, and carbon quotas. They cannot characterize the dynamic impact of behavioral events on energy consumption and carbon emissions, and the prediction models are relatively isolated, making it difficult to achieve synergistic optimization of safety, economy, and low-carbon goals.

[0025] Therefore, addressing the aforementioned problems of existing methods' inability to achieve unified fusion and behavior-driven analysis of multi-source heterogeneous data, their inability to support multi-level collaborative optimization decision-making, and their lack of dynamic adaptability based on carbon quota status, this paper addresses these issues through steps S1-S6: collecting and preprocessing historical and real-time multi-source working data, and extracting behavioral event feature vectors. This resolves the limitations of existing methods in achieving unified fusion and behavior-driven analysis of multi-source heterogeneous data. A multi-level energy consumption and carbon emission baseline dataset is established, and multi-time-domain energy consumption and carbon emission predictions are performed, outputting the prediction results. A reinforcement learning algorithm is used to output energy-saving and carbon optimization strategies, which are then validated in a simulation model. The validated energy-saving and carbon optimization strategies are then deployed and executed, and the strategy generation process is optimized based on execution feedback. This resolves the problems of existing methods' inability to support multi-level collaborative optimization decision-making and their lack of dynamic adaptability based on carbon quota status.

[0026] S1: Collect historical multi-source working data and real-time multi-source working data and preprocess them to output a multi-layer multi-source aligned dataset; S2: Extract behavioral event feature vectors from multi-layer, multi-source aligned datasets; S3: Establish a multi-level baseline dataset for energy consumption and carbon emissions based on behavioral event feature vectors and graph neural networks; S4: Based on the multi-level energy consumption and carbon emission baseline dataset, perform multi-time domain energy consumption and carbon emission prediction, and output the multi-time domain energy consumption and carbon emission prediction results. S5: Based on the multi-time-domain energy consumption and carbon emission prediction results, a reinforcement learning algorithm is used to output energy-saving and carbon optimization strategies and verify them in the simulation model; S6: Deploy the verified energy-saving and carbon optimization strategies for execution and optimize the strategy generation process based on the execution feedback.

[0027] Example 2, an embodiment of the present invention, provides a power consumption analysis method based on multi-source data fusion, based on the previous embodiment, including: In step S1, historical multi-source working data and real-time multi-source working data are collected and preprocessed to output a multi-layer multi-source aligned dataset, including the following steps A1-A3: A1: Retrieve historical and real-time multi-source working data from the historical database and online acquisition module. Both historical and real-time multi-source working data include power operation data, production process data, building environment data, distributed energy and energy storage data, price and carbon signal data, and user behavior data. Perform preliminary association between the historical and real-time multi-source working data according to timestamps and device identifiers.

[0028] A2: Perform data cleaning on historical multi-source working data and real-time multi-source working data, remove data records with severe missing values ​​and obvious anomalies, perform missing value filling, consistency verification and unit conversion on the remaining data, and unify it to the preset sampling period.

[0029] A3: Based on timestamp alignment, electrical topology mapping, and spatial location mapping, power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data are spatiotemporally aligned, and time labels, spatial region labels, and hierarchical labels are added to each data point to form a multi-layer, multi-source aligned dataset.

[0030] In step S2, behavioral event feature vectors are extracted based on the multi-layer, multi-source aligned dataset, including the following steps B1-B3: B1: Identify behavioral events in multi-layer, multi-source aligned datasets to form behavioral event samples.

[0031] Time segments related to user behavior data and production process data are selected from the multi-layer multi-source aligned dataset to identify behavioral events, including shift changes, work order changes, concentrated commuting and charging peaks. Based on the occurrence time of the behavioral events, power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data and user behavior data of a preset duration before and after the events are extracted from the multi-layer multi-source aligned dataset to form behavioral event samples.

[0032] B2: Construct behavioral event feature vectors based on behavioral event samples to form a feature library.

[0033] For each behavioral event sample, the power change magnitude and rate of change in the power operation data of each device and each area before and after the event are calculated. The comfort deviation index in the building environment data is calculated. The carbon emission factor in the price carbon signal data is used to convert the power operation data into carbon emission changes. Combined with the indicators representing process quality and cycle time in the production process data and the output changes in the distributed energy and energy storage data, a behavioral event feature vector is constructed to form a feature library.

[0034] B3: Establish a causal prior matrix based on the behavioral event feature vectors in the feature library.

[0035] Based on the statistical relationship between the feature vectors of behavioral events in the feature library and the corresponding power operation data, building environment data and carbon emission changes, correlation coefficients and mutual information indices are calculated to distinguish between rigid loads and adjustable loads, and a causal prior matrix between behavioral types, operating condition characteristics and energy consumption and carbon emission responses is established.

[0036] In this embodiment of the application, the specific steps for constructing behavioral event feature vectors and forming a feature library in step B2 are as follows: For the multi-layer, multi-source aligned dataset obtained in step S1, behavioral events are first identified based on user behavior data and production process data, including shift change events, work order insertion events, concentrated off-duty events, and charging peak events. For each type of behavioral event, the time of occurrence is recorded, and power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data within two time windows are extracted from the multi-layer, multi-source aligned dataset to form behavioral event samples.

[0037] For each behavioral event sample, the following metrics are calculated: The magnitude and rate of power change are defined as follows: at the equipment and regional levels, the active power in the power operation data is averaged according to a window to obtain the average power before the event and the average power after the event. The magnitude of power change is defined as the difference between the average power after the event and the average power before the event. The rate of change can be calculated by selecting the power values ​​of two adjacent moments before and after the event, and its value is the ratio of the power difference between the two moments to the time difference.

[0038] The comfort deviation index is calculated by selecting the indoor temperature and preset temperature of each area in the building environment data, calculating the root mean square of the temperature deviation within the entire event window, summing the squares of the differences between the indoor temperature and the preset temperature at each sampling time, taking the average and then taking the square root, which is used as the comfort deviation index of that area under this behavioral event.

[0039] The change in carbon emissions is calculated by multiplying the power in the equipment-level and regional-level power operation data by the carbon emission factor for the corresponding time period and by the sampling time interval, based on the real-time carbon emission factor recorded in the price carbon signal data. The results of each time period within the event window are then summed to obtain the change in carbon emissions caused by the event.

[0040] Operating conditions and quality characteristics involve statistically analyzing process parameters, cycle time indicators, and quality indicators related to the current behavioral event in the production process data. The mean and standard deviation changes of process parameters before and after the event are calculated, as well as the changes in pass rate or scrap rate.

[0041] The aforementioned power change magnitude, power change rate, comfort deviation index, carbon emission change, and operating condition-quality related changes are combined in a preset order to form a behavior-operating condition-energy consumption event feature vector, which is then mapped one-to-one with the corresponding behavior event type and stored in a feature library. In this way, each event sample in the feature library possesses quantifiable energy consumption and carbon emission response information, providing a foundation for subsequent construction of causal prior matrices and graph neural network modeling.

[0042] In one alternative approach, the construction of the behavioral event feature vectors and the formation of the feature library in step B2 can also employ a method based on statistical feature engineering and principal component analysis. For the identified behavioral events, a fixed-length time window is extracted from the time series data. Multiple statistical features within the window are manually designed and calculated to form a high-dimensional original feature vector. To remove redundancy and compress dimensionality, principal component analysis is used to reduce the dimensionality of the original feature vector, and the resulting low-dimensional principal components are stored in the feature library as behavioral event feature vectors.

[0043] In another alternative approach, the construction of the behavioral event feature vector and the formation of the feature library in step B2 can also employ a method based on dynamic time warping and pattern matching. Typical power or energy consumption curves for various behavioral events are extracted from historical data as reference templates. For newly identified behavioral events, the dynamic time warping distance between their time series data (e.g., the total power curve) and each reference template is calculated. The warped distance sequences from the event to all templates are combined into a feature vector to quantify its similarity and differences with various typical patterns, and then stored in the feature library.

[0044] It should be noted that this invention transforms the causal relationships implicit in multi-source data into structured features by identifying and quantifying key behavioral events such as shift changes and charging peaks. The constructed feature library integrates multi-dimensional information such as electricity, environment, process, and behavior, and establishes a quantitative mapping relationship between behavioral patterns and energy consumption and carbon emission responses by calculating correlation coefficients and mutual information.

[0045] In step S3, a multi-level energy consumption and carbon emission baseline dataset is established based on behavioral event feature vectors combined with a graph neural network, including the following steps C1-C4: C1: Construct a hierarchical heterogeneous energy consumption map based on the multi-layer multi-source aligned dataset and feature library.

[0046] A hierarchical heterogeneous energy consumption graph is constructed based on a multi-layer, multi-source aligned dataset and feature library. The hierarchical heterogeneous energy consumption graph includes device nodes, regional nodes, park nodes, behavioral event nodes, distributed energy nodes, and price carbon signal nodes. The nodes are connected by edges through electrical connection, thermal coupling, behavioral influence, and energy constraint relationships.

[0047] C2: Maps the data in the multi-layer, multi-source aligned dataset to the features of each node in the hierarchical heterogeneous energy consumption graph, and maps the feature library and causal prior matrix to the features of edges and the initial values ​​of edge weights in the heterogeneous energy consumption graph.

[0048] The power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data and user behavior data in the multi-layer multi-source aligned dataset are mapped to the features of each node in the hierarchical heterogeneous energy consumption graph. The feature library and causal prior matrix are mapped to the features of the edges and the initial values ​​of the edge weights.

[0049] C3: Input the mapped hierarchical heterogeneous energy consumption map into the graph neural network fusion model to obtain the embedding features of each node.

[0050] The hierarchical heterogeneous energy consumption graph is input into a graph neural network fusion model. A structural attention mechanism is used to weight different connection edges, and a semantic attention mechanism is used to distinguish the importance of behavioral and physical driving information. Finally, a causal prior matrix is ​​used to constrain the update of attention weights to obtain the embedding features of each device node, area node, and campus node. C4: Based on the embedding features of each node, a multi-level energy consumption and carbon emission baseline dataset is generated under given operating conditions and behavioral patterns by combining a multi-level, multi-source aligned dataset.

[0051] Based on the embedding features of equipment nodes, regional nodes, and park nodes, power operation data and price carbon signal data in multi-layer multi-source aligned datasets are aggregated and statistically analyzed to generate equipment-level energy consumption and carbon emission baseline curves, regional-level energy consumption and carbon emission baseline curves, and park-level comprehensive energy consumption and carbon emission baseline curves under given operating conditions and behavioral patterns.

[0052] In this embodiment of the application, the specific steps for generating the multi-level energy consumption and carbon emission baseline dataset in step C4 are as follows: Equipment nodes, regional nodes, park nodes, behavioral event nodes, distributed energy nodes, and price carbon signal nodes are uniformly numbered to construct a hierarchical heterogeneous energy consumption graph with a set of nodes and edges. Node features consist of statistical characteristics (such as mean, variance, maximum, and minimum values) of power operation data, production process data, building environment data, distributed energy and storage data, price carbon signal data, and user behavior data from the multi-layer, multi-source aligned dataset within a preset time window, as well as corresponding event features from the feature library. Edge features include electrical connection impedance, thermal coupling coefficient, behavioral influence intensity, and energy constraint parameters.

[0053] In the graph neural network fusion model, the node feature matrix is ​​used as the initial node representation, and the causal prior matrix is ​​used as the prior constraint for the edge weights. For each edge, a structural attention score is calculated using the node features and edge features. The structural attention score is obtained by inputting the concatenated vector of the linearly transformed node features and the edge features into a preset scoring function, which is a feedforward network with leaky linear rectifier units. For all incoming edges of the same target node, the structural attention score is converted into structural attention weights through a normalization function.

[0054] In the semantic attention mechanism, messages related to behavioral event nodes are distinguished from messages related to electrical connection nodes and thermally coupled nodes, and different semantic attention coefficients are assigned to each type of message. Finally, for each target node, the structural attention weights and semantic attention coefficients are multiplied together to obtain the aggregation weights. The features of adjacent nodes are then weighted and summed, and the updated embedding features of the target node are obtained through a non-linear activation function. The elements in the causal prior matrix serve as adjustment coefficients for the structural attention weights; when the causal prior strength of an edge is high, the weight of that edge is increased, and when the causal prior strength of an edge is low, the weight of that edge is decreased.

[0055] After completing the node embedding updates at several layers, the embedding features of equipment nodes, regional nodes, and park nodes are combined with the corresponding power operation data and price carbon signal data in the multi-layer, multi-source aligned dataset. Historical time periods under different operating conditions and behavioral patterns are aggregated and statistically analyzed. Conditional expectation or conditional quantile methods are used to generate energy consumption baseline curves and carbon emission baseline curves at the equipment, regional, and park levels. The results are then organized into a multi-level energy consumption and carbon emission baseline dataset. This approach enables the formation of more refined baseline models for objects at different levels, taking into account behavioral events and causal priors.

[0056] In an optional implementation, the generation of multi-level energy consumption and carbon emission baseline datasets in step C4 can also employ a statistical decomposition method. Based on equipment nameplate parameters, building envelope information, and historical operation logs, mechanistic models of equipment, areas, and systems are established. Using statistical methods such as nonnegative matrix factorization, load decomposition is performed on the park's total electricity meter data or key node data, breaking down the total load curve and mapping it to different equipment or areas. Combining the output of the mechanistic model with the decomposition results, energy consumption baselines for each level are generated through calibration and fitting, and corresponding carbon emission baselines are calculated based on a fixed carbon emission factor.

[0057] In another optional implementation, the generation of multi-level energy consumption and carbon emission baseline datasets in step C4 can also employ an engineering estimation method based on equipment rated parameters and operating schedules. This involves compiling the rated power, typical load rate, and equipment ledger information for all electrical equipment within the park. Combining the equipment start-up and shutdown schedules provided by the production planning system, the operating modes specified in the process procedures, and the building shift schedules, the baseline energy consumption of each piece of equipment under theoretical full-load conditions is calculated. Empirical corrections are made based on the historical operating efficiency of the equipment, and the calculation results are aggregated level by level according to electrical topology or spatial affiliation to form theoretical energy consumption baselines for equipment, regions, and the park as a whole. These baselines are then multiplied by a unified emission factor to obtain the carbon emission baseline.

[0058] It should be noted that this invention constructs a multi-layer, multi-source aligned dataset, a feature library, and a causal prior matrix, and introduces a hierarchical heterogeneous energy consumption map and graph neural network fusion model to achieve unified modeling of power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data. It forms energy consumption baselines and carbon emission baselines that take into account operating condition differences and behavioral impacts at the equipment level, regional level, and park level, thereby improving the accuracy of energy consumption level characterization in complex scenarios.

[0059] In step S4, multi-time-domain energy consumption and carbon emission predictions are performed based on the multi-level energy consumption and carbon emission baseline dataset, and the multi-time-domain energy consumption and carbon emission prediction results are output, including the following steps D1-D3: D1: Define short-term, medium-term, and long-term forecast time domains based on the park's operational needs, and establish a correspondence between the multi-level energy consumption and carbon emission baseline datasets and the multi-level multi-source aligned datasets.

[0060] Based on the park's operational needs, short-term, medium-term, and long-term forecast time domains are defined. Correspondence is established between the equipment-level, regional-level, and park-level baseline curves in the multi-level energy consumption and carbon emission baseline dataset and the corresponding multi-level, multi-source aligned datasets.

[0061] D2: Combines data from multi-level multi-source aligned datasets and multi-level energy consumption and carbon emission baseline datasets into a multi-time-domain prediction input feature sequence.

[0062] The node embedding features, multi-level energy consumption and carbon emission baseline datasets, future weather forecast information, output prediction information from distributed energy and energy storage data, and production scheduling information from user behavior data and production process data are combined into a multi-time domain prediction input feature sequence.

[0063] D3: Establish a graph-driven time series prediction model, input multi-time-domain prediction feature sequences, and output multi-time-domain energy consumption and carbon emission prediction results.

[0064] Based on a graph-driven time-series prediction model, this paper models the multi-time-domain prediction input feature sequences and outputs energy consumption and carbon emission prediction curves at the device level, the regional level, and the park level in the short, medium, and long-term prediction domains. Based on the deviation between the prediction results and the multi-level energy consumption and carbon emission baseline datasets, the paper calculates the risks of transformer load rate exceeding limits, indoor comfort exceeding limits, process quality risks, and carbon quota exceeding limits, forming a multi-time-domain energy consumption and carbon emission prediction result set.

[0065] In this embodiment of the application, the specific steps for multi-time-domain energy consumption and carbon emission prediction in step S4 are as follows: Based on the park's operational needs, the forecast time domain is divided into short-term, medium-term, and long-term forecast time domains, corresponding to forecast ranges for the next few hours, the next day, and the next week, respectively. For each level of object (equipment, region, and park), the baseline curve for the most recent period is extracted from the multi-level energy consumption and carbon emission baseline dataset. This curve, along with power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data from the multi-level multi-source aligned dataset, forms a fixed-length historical input sequence.

[0066] The historical input sequence is combined with weather forecasts for future time periods, output predictions from distributed energy and energy storage data, and production scheduling information from production process data and user behavior data to form a multi-time-domain prediction input feature sequence. This multi-time-domain prediction input feature sequence is then input into a graph-structure-driven time-series prediction model. This model includes a module for initializing node embeddings based on a hierarchical heterogeneous energy consumption graph and a module for sequence modeling in the time dimension. The sequence modeling module can employ a gated recurrent unit network (GRN) or long short-term memory (LSTM) network structure to recursively update the input features and node embeddings at each time step, outputting the predicted energy consumption and carbon emissions values ​​for each time point.

[0067] During the training phase, the predicted energy consumption and carbon emissions are compared with the actual energy consumption and carbon emissions data for the corresponding time periods. The prediction error for each level of object in each prediction time domain is calculated, and the weighted mean square error index is used as the loss function to update the parameters of the graph structure-driven time series prediction model. Different weights can be set in the loss function according to the importance of the level and the different prediction time domains. For example, higher weights can be set for the park level and the short-term prediction time domain to improve the accuracy of short-term overall load and carbon emission predictions. The multi-time domain energy consumption and carbon emission prediction result set obtained in this way not only includes power and carbon emission curves, but also, based on the deviation of the predicted values ​​from the multi-level energy consumption and carbon emission baseline datasets, can deduce the risks of transformer load rate exceeding limits, indoor comfort exceeding limits, and carbon quota exceeding limits, providing multi-dimensional constraint information for subsequent strategy optimization.

[0068] In an optional implementation, step S4, which involves multi-time-domain energy consumption and carbon emission prediction, can also employ a combined approach based on classical time series analysis and regression. Seasonally differentiated autoregressive moving average models are constructed for baseline predictions of independent historical energy consumption time series data at the equipment, regional, and park levels. External factors such as weather forecast temperatures and holiday markers are introduced into the model as regression terms to correct for weather-sensitive loads. The independent prediction results at each level are then simply summarized and multiplied by the planned unified or time-varying carbon emission factor to obtain separate energy consumption and carbon emission prediction curves.

[0069] In another optional implementation, step S4, which involves multi-time-domain energy consumption and carbon emission prediction, can also employ an engineering prediction method based on load decomposition and factor superposition. The historical total load curve is decomposed into components such as basic load, weather-sensitive load, and production plan load. Trend extrapolation is used for the basic load, a degree-day model is used for the weather-sensitive load, and the production plan load is directly calculated based on production schedules and unit consumption quotas. The prediction results of each component are superimposed over time to form the total energy consumption prediction, which is then multiplied by a carbon emission factor to obtain the carbon emission prediction.

[0070] It should be noted that this invention is based on a graph structure-driven time-series prediction model, which combines node embedding features with future weather forecasts, distributed energy output predictions, and production scheduling information to achieve short-term, medium-term, and long-term multi-time-domain energy consumption and carbon emission predictions. It can identify in advance transformer load rate exceeding limits, indoor comfort deviations, process quality risks, and carbon quota exceeding risks.

[0071] In step S5, based on the multi-time-domain energy consumption and carbon emission prediction results, a reinforcement learning algorithm is used to output energy-saving and carbon optimization strategies and verify them in the simulation model, including the following steps E1-E5: E1: Establish an expert rule base based on safety procedures and operation and maintenance experience, and transform the preset rules into action constraints and penalty parameters for energy-saving and carbon optimization strategies.

[0072] Based on safety procedures, process specifications, equipment operation manuals, and operation and maintenance experience, an expert rule base is built, and the preset rules are transformed into action constraints and penalty parameters for energy-saving and carbon optimization strategies. The preset rules include non-adjustable loads, mandatory comfort zones, safe operating ranges for energy storage, and equipment start-up and shutdown frequency limits.

[0073] E2: The action space is defined by using the multi-time domain energy consumption and carbon emission prediction results, multi-level energy consumption and carbon emission baseline datasets, and feature library as the state inputs of the reinforcement learning algorithm.

[0074] The multi-time-domain energy consumption and carbon emission prediction result set, multi-level energy consumption and carbon emission baseline dataset, and feature library are used as the state input of the reinforcement learning agent. An action space is defined, which includes adjusting production scheduling, adjusting building environment settings, adjusting the output curve of distributed energy and energy storage, and adjusting charging power. A reward function is constructed, which weights electricity cost, carbon emission, process quality risk, comfort deviation, and equipment life loss according to weights and superimposes penalty terms from the expert rule base.

[0075] E3: Establish a reward function based on the penalty parameters.

[0076] E4: The reinforcement learning algorithm searches the policy space to generate a candidate set of energy-saving and carbon optimization strategies that meet the constraints of the expert rule base and the preset reward value range.

[0077] During training and operation, the reinforcement learning algorithm selects a set of energy-saving and carbon-optimization actions based on the current state. These actions are then applied to a multi-time-domain energy consumption and carbon emission prediction result set, resulting in adjusted energy consumption and carbon emission prediction curves. The corresponding reward value is calculated using a reward function. While satisfying the constraints of the expert rule base, strategies with higher reward values ​​are retained as candidate energy-saving and carbon-optimization strategies.

[0078] E5: Input the candidate set of energy-saving and carbon optimization strategies into the simulation model for verification.

[0079] The simulation model is a coupled digital twin model of electrothermal behavior. The implementation steps are as follows: Based on the multi-layer multi-source aligned dataset and the multi-level energy consumption and carbon emission baseline dataset, a coupled simulation structure of the park power grid, building environment and production process is constructed. Each strategy in the candidate set of energy saving and carbon optimization strategies is regarded as a perturbation to the current baseline operating condition. The perturbation is applied to the equipment level, regional level and park level nodes. The power operation data, building environment data, distributed energy and energy storage data and carbon emission changes after the strategy implementation are simulated. The energy consumption and carbon emission indicators and risk indicators before and after the strategy implementation are compared. Strategies that violate the constraints of the expert rule base or cause the risk indicators to exceed the limit are eliminated. The remaining strategies are sorted according to the comprehensive benefit score and the energy saving and carbon optimization strategies are output.

[0080] The electrothermal behavior coupled digital twin model comprises three parts: a campus power grid simulation sub-model, a building thermal environment simulation sub-model, and a behavior impact simulation sub-model. The campus power grid simulation sub-model is based on the campus electrical topology, treating substations, feeders, distribution transformers, switches, and major electrical equipment as network nodes, and lines and transformers as network branches. Node power and branch power are calibrated based on voltage, current, and power measurements from power operation data, and the simulation process ensures that the injected power at each node equals the sum of the branch power and the load power.

[0081] The building thermal environment simulation sub-model abstracts each building area into a first-order thermal inertial unit with heat capacity and heat transfer coefficient. It estimates the thermal parameters of each area using indoor temperature, outdoor temperature and solar radiation data from the building environment data. It takes the cooling or heating power of the air conditioning system, the heat generation of internal equipment and the heat generation of people as heat sources, and the heat transfer of the building envelope and air infiltration as heat dissipation. It establishes a thermal balance differential equation and iteratively updates the area temperature at a given time step.

[0082] The behavioral impact simulation sub-model estimates the number of people in each area based on personnel entry and exit records and location statistics from user behavior data. It then multiplies the number of people by the per capita calorific value and per capita carbon dioxide production rate to obtain the impact of behavior on the building's thermal environment and air quality. Furthermore, based on event characteristics in a feature library, the behavioral impact simulation sub-model maps specific behavioral events to disturbances in power load and process status. For example, it increases the load on corresponding charging piles when peak charging events occur, and adjusts the load on relevant production lines when shift change events occur.

[0083] When validating the candidate set of energy-saving and carbon optimization strategies, the current multi-layer multi-source aligned dataset and multi-level energy consumption and carbon emission baseline dataset are used as the initial state of the digital twin model. The control actions in each candidate set are considered as perturbation inputs to the electricity-heat-behavior system, driving the campus power grid simulation sub-model, building thermal environment simulation sub-model, and behavior impact simulation sub-model to run in the prediction time domain, obtaining power operation data, building environment data, and carbon emission data after strategy implementation. The energy consumption and carbon emission curves after strategy implementation are compared with the baseline curves to calculate energy-saving and carbon emission reduction benefits. It is also checked for transformer overload, exceeding indoor comfort limits, and exceeding process quality risk limits. Strategies that do not meet safety constraints or lead to excessive risks are eliminated. Strategies that meet the constraints are ranked according to their comprehensive energy-saving and carbon emission reduction effects, and the strategy with the best comprehensive effect is selected as the final energy-saving and carbon optimization strategy.

[0084] In this embodiment of the application, the specific steps for outputting the energy-saving and carbon optimization strategy in step S5 are as follows: Based on safety regulations and process specifications, uninterrupted processes, critical process equipment, special environmental areas, and safe operating zones of energy storage systems are identified and compiled into an expert rule base. Each rule in the expert rule base defines triggering conditions and violation penalty coefficients. For example, it stipulates that the power of a certain type of non-adjustable load cannot be lower than a preset threshold, or that the room temperature of a certain type of area cannot exceed upper or lower limits during operating hours.

[0085] The multi-time-domain energy consumption and carbon emission prediction result set, the multi-level energy consumption and carbon emission baseline dataset, and the feature library are mapped to the state vector of the reinforcement learning agent. The state vector contains the predicted load, predicted carbon emission, predicted comfort index, and current carbon quota status information for key equipment and areas. The action space is defined as the set of policy adjustments that can be executed within a control cycle, specifically including: actions to fine-tune the shift schedule of production, actions to fine-tune the temperature and ventilation strategies of the building environment, actions to adjust the charging and discharging power curves of distributed energy and energy storage, and actions to adjust the power limit of charging piles, etc.

[0086] In the reward function design, electricity cost, carbon emissions, comfort deviation, process quality risk, and equipment lifespan loss are defined as five cost components. Time-of-use electricity prices and carbon emission factors from price carbon signal data are used to convert predicted energy consumption and carbon emissions into corresponding electricity and carbon emission costs. Comfort deviation is calculated using temperature and air quality indicators from building environment data. Process quality risk is constructed using cycle time and quality indicators from production process data. Equipment lifespan loss is estimated using equipment start-up / shutdown frequency and overload level. The five cost components are weighted and summed according to preset weights, and fixed or dynamic penalty coefficients are added to actions that violate rules in the expert rule base to form the total cost. The negative of the total cost is used as the reward value for the reinforcement learning agent in a certain control cycle, causing the agent to tend to choose action sequences that reduce the total cost.

[0087] In an optional implementation, the energy-saving and carbon optimization strategy output in step S5 can also employ a strategy generation method based on an expert experience rule base and logical judgment. According to equipment operating procedures, process constraints, and safety standards, a set of "if-then" rules, including priority order and threshold judgment, is manually established. The system monitors energy consumption, power, and environmental parameters in real time. When the monitored values ​​trigger specific rule conditions, the corresponding preset control commands are automatically invoked, such as adjusting the air conditioning temperature setpoint, shutting down non-critical equipment, or switching the energy storage charging and discharging mode.

[0088] In another optional implementation, the energy-saving and carbon optimization strategy output in step S5 can also adopt a static scheduling method based on linear programming or multi-objective optimization. This method uses electricity cost and carbon emissions as optimization objectives, and equipment operating limits, comfort ranges, and process constraints as boundary conditions to establish a mixed-integer linear programming model of the park's energy flow. Based on predicted load, electricity price, and renewable energy output data, this model is solved daily to generate a schedule for equipment start-up and shutdown, energy storage charging and discharging, and load transfer for each time period of the following day, serving as the optimization strategy.

[0089] It should be noted that this invention couples the expert rule base with reinforcement learning strategy search, and uses a coupled digital twin model to perform perturbation simulation verification on the candidate set of energy-saving and carbon optimization strategies. Under the premise of meeting safety procedures and process specifications, strategies with excellent energy-saving effect, carbon emission reduction effect and operation risk indicators are selected, realizing multi-level and multi-objective collaborative optimization from equipment level to park level.

[0090] In step S6, the verified energy-saving and carbon optimization strategy is deployed for execution, and the optimization strategy generation process is performed based on the execution feedback, including the following steps F1-F2: F1: The energy-saving and carbon optimization strategies output by the digital twin model coupled with electrothermal behavior are sent to the park control terminal to jointly control the production process control system, building environment control system, distributed energy and energy storage control system, and charging control system. During the strategy execution process, power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data are collected to form a strategy execution feedback dataset.

[0091] F2: Compare the policy execution feedback dataset with the multi-level energy consumption and carbon emission baseline dataset and the multi-time domain energy consumption and carbon emission prediction result set, calculate the deviation between the actual energy saving benefit and the actual carbon emission reduction benefit and the predicted value, update the node features and edge weights in the hierarchical heterogeneous energy consumption graph according to the source of the deviation, update the parameters of the graph neural network fusion model and the graph structure-driven time series prediction model, and use the policy execution feedback dataset as the experience replay data of the reinforcement learning algorithm to incrementally train the energy saving and carbon optimization strategy to form a policy effect profile, which evolves over time.

[0092] Step F2 includes: calculating the carbon quota consumption and carbon quota remaining amount for each time period in the strategy execution feedback dataset based on the real-time carbon emission factor and carbon quota surplus information recorded in the price carbon signal data, and constructing the carbon quota state curve.

[0093] When the remaining amount of carbon quota, as represented by the carbon quota state curve, is lower than a preset threshold, the weight of carbon emissions in the reward function is increased to a preset percentage and the weight of electricity cost is decreased to a preset percentage.

[0094] When the remaining carbon allowance, as represented by the carbon allowance state curve, is higher than a preset threshold, the weight of carbon emissions in the reward function is reduced to a preset percentage, and the weight of electricity costs is increased to a preset percentage.

[0095] The carbon quota state curve and weight adaptive adjustment can be achieved as follows: Based on the total annual or assessment cycle carbon allowance recorded in the price carbon signal data, and the carbon allowance consumption records for each time period, the remaining carbon allowance is defined as the current total carbon allowance minus the cumulative carbon emissions up to the current moment. The carbon emissions for each control cycle are obtained by multiplying and summing the power operation data and the real-time carbon emission factor, and then subtracted from the remaining carbon allowance to obtain a carbon allowance status curve over time. The carbon allowance status curve reflects the degree of matching between the current carbon allowance consumption progress and the remaining time of the assessment cycle.

[0096] In the reward function, the weights corresponding to electricity costs and carbon emissions are set as adaptive weights that can change with the carbon quota status. When the carbon quota status curve indicates that the remaining carbon quota is lower than the preset carbon quota safety threshold, and the predicted energy consumption trend may lead to the depletion of carbon quota before the end of the assessment period, the weight corresponding to carbon emissions is increased from the base value to a higher level, and the weight corresponding to electricity costs is reduced accordingly. This makes the reinforcement learning agent pay more attention to carbon emission constraints in the subsequent policy search process, and prioritize energy-saving and carbon optimization strategies with better carbon emission reduction effects.

[0097] When the carbon quota status curve indicates that the remaining carbon quota is higher than the preset carbon quota safety threshold, and the predicted energy consumption trend will not lead to carbon quota overruns, the weight corresponding to carbon emissions is gradually reduced from a higher level to a base value, while the weight corresponding to electricity costs is increased accordingly. This ensures that the energy-saving and carbon optimization strategy, while maintaining carbon quota constraints, is biased towards a more cost-effective approach. In one specific implementation, the remaining carbon quota can be divided into three zones: a tight zone, a normal zone, and a slack zone. The highest carbon weight is used in the tight zone, a medium carbon weight in the normal zone, and the lowest carbon weight in the slack zone. A smooth transition between these zones is achieved through linear interpolation or a piecewise function. This adaptive weight adjustment mechanism allows the energy-saving and carbon optimization strategy of this invention to dynamically adjust its optimization objectives based on the carbon quota consumption progress, maintaining the rationality and executability of the strategy under different carbon constraint scenarios.

[0098] In summary, this invention achieves a quantitative representation of dynamic coupling relationships in complex energy consumption scenarios through multi-source data fusion with a unified spatiotemporal benchmark and behavioral event feature extraction. By using hierarchical heterogeneous graph neural network modeling, it solves the problem of traditional methods struggling to depict multi-level state coordination from equipment to the entire industrial park. Through graph-structured time-series prediction incorporating external information and multi-dimensional risk assessment, it achieves the identification of operational risks. Through reinforcement learning constrained by expert rules and digital twin closed-loop verification, it generates an executable optimization strategy that coordinates safety, economy, and low-carbon goals, and dynamically adjusts it based on carbon quotas and feedback, overcoming the problems of existing technologies having a single optimization objective and lacking adaptability.

[0099] Example 3 is an embodiment of the present invention, which provides a power consumption analysis system based on multi-source data fusion, including: a data acquisition module, a feature extraction module, a baseline establishment module, a prediction module, a strategy output module, and a strategy optimization module; The acquisition module collects historical multi-source working data and real-time multi-source working data, performs preprocessing, and outputs a multi-layer multi-source aligned dataset. The feature extraction module extracts behavioral event feature vectors based on a multi-layer, multi-source aligned dataset. The baseline establishment module establishes a multi-level energy consumption and carbon emission baseline dataset based on behavioral event feature vectors combined with graph neural networks. The prediction module performs multi-time-domain energy consumption and carbon emission prediction based on multi-level energy consumption and carbon emission baseline datasets, and outputs multi-time-domain energy consumption and carbon emission prediction results. The strategy output module outputs energy-saving and carbon optimization strategies based on multi-time-domain energy consumption and carbon emission prediction results using reinforcement learning algorithms, and verifies them in a simulation model. The strategy optimization module is the process of issuing and executing the verified energy-saving and carbon optimization strategies and optimizing the strategy generation process based on the execution feedback.

[0100] This embodiment also provides an electronic device applicable to a power consumption analysis method based on multi-source data fusion, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the power consumption analysis method based on multi-source data fusion as proposed in the above embodiment.

[0101] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a power consumption analysis method based on multi-source data fusion as proposed in the above embodiment.

[0102] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing power consumption analysis based on multi-source data fusion proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0103] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A power consumption analysis method based on multi-source data fusion, characterized in that: include, Collect historical and real-time multi-source working data and preprocess them to output a multi-layer multi-source aligned dataset. Extract behavioral event feature vectors from multi-layer, multi-source aligned datasets; A multi-level baseline dataset for energy consumption and carbon emissions is established by combining behavioral event feature vectors with graph neural networks. Based on the multi-level energy consumption and carbon emission baseline dataset, multi-time domain energy consumption and carbon emission predictions are performed, and the multi-time domain energy consumption and carbon emission prediction results are output. Based on the multi-time-domain energy consumption and carbon emission prediction results, a reinforcement learning algorithm is used to output energy-saving and carbon optimization strategies and verify them in a simulation model. The validated energy-saving and carbon optimization strategies are deployed for execution, and the strategy generation process is optimized based on the execution feedback.

2. The power consumption analysis method based on multi-source data fusion as described in claim 1, characterized in that: The step of extracting behavioral event feature vectors from a multi-layer, multi-source aligned dataset includes: Identify behavioral events in multi-layer, multi-source aligned datasets to form behavioral event samples; Based on behavioral event samples, construct behavioral event feature vectors to form a feature library; A causal prior matrix is ​​established based on the behavioral event feature vectors in the feature library.

3. The power consumption analysis method based on multi-source data fusion as described in claim 2, characterized in that: The process of establishing a multi-level energy consumption and carbon emission baseline dataset based on behavioral event feature vectors combined with graph neural networks includes: A hierarchical heterogeneous energy consumption map is constructed based on a multi-layer, multi-source aligned dataset and feature library. The data in the multi-layer multi-source aligned dataset is mapped to the features of each node in the hierarchical heterogeneous energy consumption graph, and the feature library and causal prior matrix are mapped to the features of the edges and the initial values ​​of the edge weights in the heterogeneous energy consumption graph. The mapped hierarchical heterogeneous energy consumption map is input into the graph neural network fusion model to obtain the embedding features of each node; Based on the embedding features of each node, a multi-level energy consumption and carbon emission baseline dataset is generated under given operating conditions and behavioral patterns by combining a multi-level, multi-source aligned dataset.

4. The power consumption analysis method based on multi-source data fusion as described in claim 3, characterized in that: The method involves predicting energy consumption and carbon emissions in multiple time domains based on multi-level energy consumption and carbon emission baseline datasets, and outputting the prediction results, including: Based on the park's operational needs, short-term, medium-term, and long-term forecast time domains are defined, and a correspondence is established between the multi-level energy consumption and carbon emission baseline datasets and the multi-level, multi-source aligned datasets. The data from the multi-layer, multi-source aligned dataset and the multi-level energy consumption and carbon emission baseline dataset are combined into a multi-time-domain prediction input feature sequence. A graph-driven time-series prediction model is established, which takes multi-time-domain prediction feature sequences as input and outputs multi-time-domain energy consumption and carbon emission prediction results.

5. The power consumption analysis method based on multi-source data fusion as described in claim 4, characterized in that: The process of using reinforcement learning algorithms to output energy-saving and carbon optimization strategies based on multi-time-domain energy consumption and carbon emission prediction results, and then validating them in a simulation model, includes: An expert rule base was established based on safety procedures and operation and maintenance experience, and the preset rules were transformed into action constraints and penalty parameters for energy-saving and carbon optimization strategies. The action space is defined by using multi-time-domain energy consumption and carbon emission prediction results, multi-level energy consumption and carbon emission baseline datasets, and feature libraries as state inputs to the reinforcement learning algorithm. Establish a reward function based on the penalty parameters; The reinforcement learning algorithm searches the policy space to generate a candidate set of energy-saving and carbon optimization strategies that meet the constraints of the expert rule base and the preset reward value range. The candidate set of energy-saving and carbon optimization strategies is input into the simulation model for verification.

6. The power consumption analysis method based on multi-source data fusion as described in claim 5, characterized in that: The step of inputting the candidate set of energy-saving and carbon optimization strategies into the simulation model for verification includes: Input the candidate set of energy-saving and carbon optimization strategies into the digital twin model of electrothermal behavior; The electrothermal behavior coupled digital twin model constructs a coupled simulation structure of the park power grid, building environment, and production process based on a multi-layer multi-source aligned dataset and a multi-level energy consumption and carbon emission baseline dataset. Each strategy in the candidate set of energy-saving and carbon optimization strategies is regarded as a perturbation to the current baseline operating conditions. The perturbation is applied to equipment-level, regional-level, and park-level nodes to simulate power operation data, building environment data, distributed energy and energy storage data, and carbon emission changes after strategy implementation. The model compares energy consumption and carbon emission indicators and risk indicators before and after strategy implementation. Strategies that violate expert rule base constraints or cause risk indicators to exceed limits are eliminated. The remaining strategies are sorted according to their comprehensive benefit scores, and the energy-saving and carbon optimization strategies are output.

7. The power consumption analysis method based on multi-source data fusion as described in claim 6, characterized in that: The process of issuing and executing the verified energy-saving and carbon optimization strategies and generating optimization strategies based on execution feedback includes: The energy-saving and carbon optimization strategies output by the electrothermal behavior coupled digital twin model are sent to the park control terminal to jointly control the production process control system, building environment control system, distributed energy and energy storage control system, and charging control system. During the strategy execution process, power operation data, production process data, building environment data, distributed energy and energy storage data, price carbon signal data, and user behavior data are collected to form a strategy execution feedback dataset. The strategy execution feedback dataset is compared with the multi-level energy consumption and carbon emission baseline dataset and the multi-time domain energy consumption and carbon emission prediction result set. The deviation between the actual energy saving benefit and the actual carbon emission reduction benefit and the predicted value is calculated. According to the source of the deviation, the node features and edge weights in the hierarchical heterogeneous energy consumption graph are updated respectively. The parameters of the graph neural network fusion model and the graph structure-driven time series prediction model are updated. The strategy execution feedback dataset is used as the experience replay data of the reinforcement learning algorithm to incrementally train the energy saving and carbon optimization strategy, forming a strategy effect profile. The strategy effect profile evolves over time.

8. A power consumption analysis system based on multi-source data fusion, employing the power consumption analysis method based on multi-source data fusion as described in any one of claims 1-7, characterized in that, include: The system includes a data acquisition module, a feature extraction module, a baseline establishment module, a prediction module, a policy output module, and a policy optimization module. The acquisition module collects historical multi-source working data and real-time multi-source working data, performs preprocessing, and outputs a multi-layer multi-source aligned dataset. The feature extraction module extracts behavioral event feature vectors based on a multi-layer, multi-source aligned dataset. The baseline establishment module establishes a multi-level energy consumption and carbon emission baseline dataset based on behavioral event feature vectors combined with graph neural networks. The prediction module performs multi-time-domain energy consumption and carbon emission prediction based on multi-level energy consumption and carbon emission baseline datasets, and outputs multi-time-domain energy consumption and carbon emission prediction results. The strategy output module outputs energy-saving and carbon optimization strategies based on multi-time-domain energy consumption and carbon emission prediction results using reinforcement learning algorithms, and verifies them in a simulation model. The strategy optimization module is the process of issuing and executing the verified energy-saving and carbon optimization strategies and optimizing the strategy generation process based on the execution feedback.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power consumption analysis method based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power consumption analysis method based on multi-source data fusion as described in any one of claims 1 to 7.

Images